client GC: separate allocrunner from allocdir management

[tgross]

After an allocation is terminal, the Nomad client needs to keep the allocation directory around so that users can read their logs to debug failed allocations. But to do so, Nomad keeps the entire AllocRunner alive, which uses up memory resources we should be freeing for the user's workloads (sometimes substantially, ex. #25269), and complicates doing the actual GC (ex. #25123). At the end of the day users want the alloc dir to persist as long as they don't otherwise need the disk space, but shouldn't have to pay for all this overhead.

Nomad should instead split the allocrunner and the alloc directory into separate objects, so that we can abandon the allocrunner once the allocation is terminal, and keep the alloc directory around until GC so we can serve /v1/client/fs APIs like Stream Logs.

There are several subtasks to figure out, mostly around the cleanup operations in allocrunner hooks:

• The csi_hook has a Destroy hook method (triggered on GC). It looks like this only cancels any in-flight requests. Can we safely move this to a Postrun hook?
• The consul_hook has a Destroy hook method. This exactly copies the Postrun hook method. Why do we feel like we need both here? Do we have a resource cleanup problem that we're trying to paper over?
• The identity_hook has a Destroy hook method. This exactly copies the PreKill and Shutdown hook methods. Why do we feel like we need all three here? Why isn't there a Postrun hook?
• The allocdir_hook has a Destroy hook method which is used to cleanup on GC. We would need to move this out of the allocrunner entirely and into whatever "alloc directory proxy" we build.
• The alloc FS RPCs defined in fs_endpoint.go will need reworking to use some new client object as a proxy to the allocdir. It looks like there are a small number of methods to implement (ex. we need the allocation ID to find the allocdir and its namespace to check auth).
• There will be a bunch of refactoring to do in client/client.go and client/gc.go so that we're ensuring the allocrunner is completely dead before cleaning up the disk for GC.

(internal ref https://hashicorp.atlassian.net/browse/NET-12306)